FDA Seeks Feedback on Generative AI Medical Device Rules

The US Food and Drug Administration's Center for Devices and Radiological Health posted a discussion paper titled Considerations for the Regulation of Generative AI-Enabled Medical Devices on 18 August 2026 and opened docket FDA-2026-N-7874 for comment through 19 October 2026. The Digital Health Center of Excellence asked for feedback on 26 numbered questions and said partial responses are welcome. The paper is not draft or final guidance, does not propose or implement policy changes, and does not set the agency's evidence expectations for future marketing submissions.

What FDA is asking before it writes guidance

Rick Abramson, director of the Digital Health Center of Excellence, told STAT's Mario Aguilar in an interview published 24 August 2026 that the goal is formal policy guidance. The ecosystem is expecting clarity, he said, and can expect both broad guidance on generative AI and more narrowly constructed specialty guidance on topics of special interest or complexity. That interview is the public translation of a paper that FDA otherwise framed as discussion only. The agency's own landing page repeats a boundary manufacturers quote constantly: FDA does not regulate GenAI as such; it regulates medical devices.

The paper organizes discussion around risk assessment, premarket evaluation, postmarket monitoring, and other topics, including agentic systems. A two-axis risk picture scores how independently a function acts — from information-only through action-directing, supervised, and autonomous — against how severe the harm would be if the output is wrong. Evidence expectations scale with position on that grid, not with the mere presence of a large language model. A disclaimer such as talk to your doctor, the paper's logic holds, does not by itself make an instruction less directive.

Premarket evaluation is described as competency-based. Clinical confirmation, FDA writes, might not require a prospective clinical study in every case. Five approaches of increasing rigor run from retrospective evaluation and shadow deployment through standardized patient interactions, clinician adjudication, and prospective studies. Because generative systems can behave unpredictably after launch, the paper also asks whether stronger postmarket monitoring should compensate for premarket tests that cannot exhaust every prompt path. That question is why comments on this docket will shape later guidance even though the paper itself changes no rule.

How agentic medical software changes the device question

Agentic AI — systems that plan multi-step work, use tools, and take actions — gets its own attention. Some agentic functions could themselves be medical devices if they control another device or make real-world clinical decisions. Question 26 asks what extra premarket and postmarket evaluation those systems need beyond non-agentic generative devices, including how autonomous multi-step action, tool use, and reduced human review should show up in acceptance criteria. Manufacturers selling into the United States, including firms headquartered abroad, now have a docket number and a date rather than a rumor about forthcoming AI device rules.

The paper grows out of a November 2024 Digital Health Advisory Committee public meeting on total product life cycle issues for generative AI devices. STAT's Abramson interview and the 18 August paper together tell industry to expect both a broad generative AI policy document and narrower specialty pieces, while reminding everyone that this month's PDF is still only a request for feedback. Related PromptCrates coverage of Anthropic Claude protein binders and hit rate, Google Gemini Enterprise for legal and financial services, and the OpenAI, Anthropic, and Google rogue-AI defense letter shows how lab science, enterprise agents, and frontier-model governance are landing in the same season as FDA's device questions.

What device makers should put on the 19 October docket

FDA invited device manufacturers, clinicians, researchers, and the public to comment. Interested parties do not have to answer all 26 questions. The practical reading for sponsors is to place each generative function on the independence-versus-harm grid, say which of the five confirmation rungs they think is proportionate, and address whether agentic tool-use needs extra checkpoints before irreversible actions. Because the paper is not guidance, a comment cannot be dismissed as fighting a final rule; it is the moment to argue about evidence burden before reviewers treat a future guidance as the default.

Global makers that sell into the US market should treat 19 October 2026 as a calendar item even if their software is built elsewhere. A function that only drafts a note is a different risk object from a function that tells a patient to go to an emergency department or that autonomously starts an order set. File this as health-science news: an 18 August 2026 discussion paper, a 24 August STAT interview with Abramson, 26 questions, docket FDA-2026-N-7874, comments due 19 October 2026, a two-axis risk frame, five clinical-confirmation approaches, and an explicit warning that agentic software can be a device if it controls hardware or makes real-world decisions.

Sources

• Considerations for the Regulation of Generative AI-Enabled Medical Devices — FDA DHCoE
• FDA digital health leader promises generative AI regulatory guidance is coming — STAT (Mario Aguilar), 24 August 2026